Environmental Engineering Reference
In-Depth Information
This description is generic and it leaves many design decisions unanswered.
For example, planners differ in how they describe the goals they are to achieve,
the environment they execute in, and their database of potential actions. Each
of these descriptions is given in a computer-based language. This language is
very important since it defines what the system is capable of doing and how
it will do it. It also defines what the system cannot do. If the language used
is too limited, it may not be able to describe some aspect of the domain,
potentially limiting the planner's ability to handle some situations.
Planners differ in the speed at which they come to decisions. Some planners
are slow and deliberative, while others are quick and reactive. In general, the
slow deliberative planners make plans that are more globally optimal and
strategic in nature. The reactive planners tend to examine the environment
and choose from a highly constrained set of plans. They are tactical in nature
and work well in rapidly changing environments where the time for slow careful
choice is not available. In many real world systems, either the planner is
designed to handle both deliberation and reactivity, or two separate planners
are integrated together.
All robust planners must deal with the failure of an action during exe-
cution. Some planners have low level strategies on hand and when a failure
occurs, they immediately attempt to repair the plan. Others have a database
of alternative ways to achieve an objective, and when an attempt fails, they
analyze the current environment and choose another plan. In domains where
the environment changes very rapidly relative to the planner decision time or
where action failure is a regular occurrence, the planner may take the possi-
bility of failure into consideration during plan creation. These systems give
preference to robust plans that can help recover from likely failures even if
the robust plan has a higher cost in resources than the alternatives.
Some planners convert higher level tasks into low level actions just before
execution. By using this strategy, they commit fewer resources to any one plan
and can quickly react to changes in the environment. Unfortunately, the plan
that is ultimately executed will often be suboptimal, particularly if two or
more tasks are competing for the same resource. Other planners map the top-
level goals into a complete series of small actions that take place in a time-
sequenced manner. The advantage of this approach is that a more globally
optimal plan can be created. Its disadvantage is that a failure occurring in
one step can cause the rest of the plan to be abandoned and re-planned.
This is computationally expensive and time consuming. Re-planning can also
cause problems if the domain looks forward into the plan and begins the
commitment of resources based on the expected plan. In these domains, the re-
planning step must use repair strategies in an attempt to maintain most of the
original plan.
Planners must be sensitive to an action's cost in resources. In computer
domains, such as software agents, actions have small costs and plan selection
can usually ignore resource issues. In other domains, like spacecraft, some
actions commit resources that cannot be replenished (such as propellant).
